Author: Bao Yonggang | Editor: Liu Wei
With CNY 50 million on the table, Dr. Zhang ultimately decided not to start his own company.
He had just left an acquired firm when he encountered a new wave of chip startups. A prominent VC was willing to provide CNY 50 million for him to assemble a team and serve as CEO. Dr. Zhang frequently met with investors and scouted for core members; the project had already reached the launch stage.
Soon after, he stopped.
After speaking with investors and others in the market, Dr. Zhang felt that his capabilities did not fully align with what entrepreneurship truly required.
In the spring of 2026, this was an anomalous decision. At that time, valuations for popular chip projects could increase tenfold in three months, and technical leads from major firms were constantly being encouraged to leave and start their own businesses. While most people worried about missing the window, Dr. Zhang chose to stay on the sidelines even when the money was right in front of him.
Wang Xiaofeng was already inside the arena.
Wang Xiaofeng, who came from a technical background, had worked at several domestic AI chip companies. The appeal of entrepreneurship for him was simple: finally, he could execute something based on his own judgment.
He was working on one of the hottest directions of 2026, LPU chips. During a funding discussion, an investor asked him: What programming model should an LPU adopt?
For GPUs, the answer is SIMT. For LPU chips, there is no ready-made answer.
Wang Xiaofeng knew clearly that this was not a problem solvable with a few pages of PPT. If he could settle down and research for two or three months, he believed he could at least produce a preliminary solution.
The problem was, he didn't have those two or three months. "There's simply no time, and I can't calm down to study slowly. Everyone on the team is under immense pressure."
It wasn't because the company lacked funds. After securing its first investment in early 2026, it completed three to four rounds of financing in over half a year, with valuation rising from initial hundreds of millions to nearly CNY 1 billion.
Dr. Zhang faced CNY 50 million and chose not to start up, while Wang Xiaofeng was pushed forward by round after round of financing.
In the 2026 chip startup boom, two clocks running at different speeds began to pull against each other: capital measured time in weeks and months, while chip R&D still measured time in years.
This boom featured more counter-intuitive phenomena: the less cash-strapped the chip company, the easier it was to raise money. By the second half of the year, capital began to cool down, yet chip talent salaries continued to double.
Why are investors willing to push a chip company's valuation up tenfold in a few months? Facing the same boom, why do some rush to start businesses while others, having secured funding, choose to give up? Add the author's WeChat BENSONEIT to discuss how long the chip boom can last.
01、The Previous GPU Listings Begin to Deliver Returns; The Next Inference Chip Startup Wave Heats Up
Sarah from Tuochi Headhunting has been in the chip industry for nearly a decade. Around October 2025, she began encountering more and more new chip companies and investment institutions. By the end of the year, her client base increased significantly.
At that time, many companies emerging in 2026 had not yet surfaced, but Sarah already sensed that the market had changed.
Similar changes appeared upstream in the supply chain.
Li Wei, who works in the chip IP business, found that starting in the second half of 2025, some projects that had been discussed for a long time but shelved due to lack of funds were restarted. Those receiving funds included not only new companies but also some established firms that had survived the troughs of 2023 and 2024.
Why did capital start investing in chips again?
A key reason is that the previous wave of AI chip investments finally allowed investors to see exits again.
The listings and high market capitalizations of Moore Threads and MetaX simultaneously stimulated both investors who had backed domestic GPUs and those who missed this opportunity. The former saw substantial paper returns and were willing to continue betting on chips, while the latter began searching for the next batch of potential IPO candidates.
But this investment boom does not belong to all chip companies.
Veteran chip investor Xue Haofei observed that in 2026, what truly heated up were AI, computing power, and the chips and infrastructure surrounding them. The heat for non-AI semiconductor assets was even weaker than in the previous two years.
While capital saw exit channels, AI inference indeed brought new opportunities for chip startups.
In the training era, a general-purpose GPU could handle most core tasks. In the inference era, latency, throughput, memory bandwidth, power consumption, and token costs became more critical. Agents, long contexts, and complex reasoning further amplified these differences, leading the industry to form a consensus: not all AI tasks necessarily need to be handled by the same type of GPU.
Even NVIDIA, the biggest winner among GPUs, began introducing specialized architectures like LPUs into its inference systems.
Chen Junjie, who is starting his second venture, summarized the previous AI chip startup wave as "domestic substitution": making a chip similar to NVIDIA's GPU and replacing NVIDIA through compatible ecosystems.
For this current wave, he judged that the capital story of domestic substitution has entered its latter half, with new opportunities shifting toward AI-specific chips and heterogeneous division of labor.
Two forces collided from late 2025 to early 2026.
On one side, the listings of companies like Moore Threads and MetaX gave capital a renewed view of exit channels for chip investments.
On the other side, new problems brought by AI inference created new imaginative space for architectures such as LPUs, 3D stacking, compute-in-memory, and near-memory computing.
One opened the capital window, and the other opened the industrial window.
After the Spring Festival, both windows opened simultaneously.
02、Valuation Multiplies 10x in 3 Months; The Less Cash-Hungry, The Easier To Raise
After the Spring Festival of 2026, the heat in the primary chip market rose rapidly.
Xue Haofei observed that the period after the Spring Festival was the fastest stage for project financing and valuation increases in this wave. Some hot projects doubled their valuations in one or two months. A quantum chip company had an angel round valuation of approximately CNY 200 million after the Spring Festival; by May, when it announced its financing news, its valuation had reached approximately CNY 2 billion, a tenfold increase in two to three months.
This speed of financing and valuation growth was uncommon during the 2021 wave of large AI chip startups.
The speed of valuation growth reflected market heat, but accompanying the rise in valuations was a seemingly counter-intuitive phenomenon: the less cash-strapped the chip company, the easier it was to raise money.
The reason is related to the long cycle of chip R&D.
A large chip requires sustained long-term investment from design, tape-out, and return, to software adaptation and customer verification. Xue Haofei estimated that some large chip projects need to reserve close to CNY 2 billion in cash upfront.
If a company's valuation is too low, it is difficult to raise enough money in one go. Investors would instead worry whether the company has the ability to survive until the product is actually built.
A loop that appears to defy common sense emerged.
Teams with strong backgrounds and capabilities are more likely to receive high valuations. High valuations allow companies to raise more cash. The more sufficient the cash reserves, the lower the probability of the company dying mid-R&D. Subsequent investors therefore perceive lower risk and are more willing to provide funds.
Rocket handles financing-related work at an LPU company. Although the company already has ample cash reserves—enough to last a long time given the current team size—he still wants to continue raising funds.
"The less I need money, the more others will give me money. The more I need money, the more others take advantage of my illness to demand my life."
This is Rocket's experience from years of doing financing.
Continuing to raise funds isn't just about inflating valuations. Experienced entrepreneurs know that capital windows are often much shorter than chip R&D cycles, so they must grab money as quickly as possible when available.
For a chip company, the real danger is not having little cash when the capital market is hottest, but needing to raise more money two years later and coinciding with a market where no investors are willing to act.
But money does not flow evenly to all chip companies.
In the first half of 2026, the chip companies investors were most willing to bet on were concentrated in a few specific directions.
03、LPU, 3D, Compute-in-Memory: Taking Over From Domestic Substitution
The 2021 chip startup boom also saw valuation rises, talent wars, and executives from major firms starting businesses. But in the two booms, the companies investors were willing to bet on were different.
Five years ago, investors had overseas reference answers. For GPUs, CPUs, EDA, analog chips, semiconductor equipment, and materials, overseas markets had proven the demand and roughly what the products should look like.
Chinese startups were mostly answering one question: Who can build it?
This was like filling in the blanks under the logic of domestic substitution.
In 2026, the questions are clear, but the answers are not.
Inference requires lower latency, higher bandwidth, and lower costs. Advanced processes and high-end storage remain constrained, and models are constantly changing. What kind of chip is actually needed?
Continuing to upgrade GPUs is one answer; LPUs are another. Compute-in-memory, near-memory computing, dedicated ASIC chips, combined with 3D stacking technology, could all become answers.
These new architectures without standard answers became the directions capital was most willing to bet on in 2026.
LPUs are the most typical example. Companies like Yuanchuan Micro, Memsilicon, Xin Gan Future, Hanxu Technology, and Fangqing Technology have generated significant attention for LPUs.
However, LPUs have not yet become a clearly defined chip category like GPUs. Different startups' visions of LPUs are not entirely identical. Startups are also willing to use the LPU label and welcome more peers joining the field.
"When the industry hasn't formed solid cognition, the emergence of competitors is, to some extent, a good thing. First establishing consensus among the market and investors benefits the entire track." Rocket's attitude is clear. "First, we must make investors believe LPU is a new track worth betting on; then it's time to compare whose team is stronger and whose product is faster."
3D chips and compute-in-memory chips share similar characteristics.
AI inference is increasingly constrained by memory bandwidth, while domestic firms face restrictions on advanced nodes, high-end packaging, and HBM. Placing storage and compute units closer together via 3D methods to solve Decode-stage bottlenecks with higher bandwidth is a technology path with genuine industrial demand.
But there is still a long distance between "technical logic holds" and "company worth billions."
At this stage, startups actively seek scenarios that major firms truly care about. If a PC, server, or OEM manufacturer prepares to launch a new product, startups consider whether their 3D chips can be embedded. Once recognized by major firms, the solution gains both customer endorsement and a capital narrative.
"Sometimes startups, investment institutions, and ideal customers 'fabricate stories' together," described a 3D chip entrepreneur.
The word "story" here is not derogatory. All early tech investing is essentially pricing a future that has not yet happened.
The real question is: How far ahead has the story run, and how closely has engineering validation followed?
No matter how many stories are told about LPUs, 3D, compute-in-memory, and high bandwidth, someone eventually must build the chip and prove in real models and customer scenarios that it indeed reduces costs and improves token output.
In the 2026 startup and financing boom, many proofs have not yet occurred, but capital has already set a very high price for these new architectures.
Once money enters these companies, it quickly flows to the next scarce resource: chip talent.
04、NPU Engineers With Two Years of Experience Command Annual Salaries of Over One Million
Capital entering chip companies quickly turns into hiring budgets.
In the summer of 2026, signs of cooling appeared on the investment side, yet HR departments and headhunters at chip companies became increasingly busy. Investment institutions deployed funds between March and June; companies started expanding teams only after receiving money. There is a natural lag of several months between the capital market and the talent market.
Salaries became a price signal as intuitive as valuations in this boom.
Senior chip headhunters at Qianheng Consulting recently encountered an impressive case: An engineer with only two years since obtaining a master's degree, having worked on NPU Cores for two years, received an annual salary exceeding CNY 1 million.
NPU R&D engineers might be the most sought-after talent in this boom.
Sarah said the reasons are direct: traditional AI chip companies want NPU talent; newly emerging LPU companies want them; embodied intelligence chip companies are also competing for them. Different architectural stories all converge into the same NPU talent pool.
Million-yuan salaries are still minority cases. A more common trend is that two years of NPU experience commands around CNY 700,000, and three years can reach CNY 900,000, significantly higher than the previous trough period.
NPU specialists are not the only ones being fought over.
As super-nodes become hotspots for computing infrastructure, related talent is equally scarce. Sarah is helping recruit for a relevant team, which targeted senior engineers with salaries over CNY 1 million.
Sarah suggested that if they really wanted to poach these people, salaries might need to double. The response was blunt: "Doubling salaries is negotiable."
Looking only at these numbers, one might recall the chip talent wars around 2021, but the two booms are not exactly the same.
Sarah remembered that salary hikes in the previous wave were more like universal markups. Candidates without fully matching experience might still get 50% or even double raises.
In 2026, chip companies are still willing to offer high prices, but they are clearly more selective.
An engineer with five or six years of experience, if neither capability nor direction is outstanding, might earn only CNY 400,000 to CNY 500,000 annually. Conversely, even with only two or three years of experience, if they happen to work on the most scarce NPU or chip interconnect roles, they can command million-yuan salaries.
In the previous wave, capital paid premiums largely for "big company backgrounds" and resumes. In this wave, more companies are beginning to pay premiums for specific capabilities that can solve immediate problems.
Sarah judges that the high premium for NPU and interconnect talent in this wave is difficult to sustain long-term.
Beyond changes in talent pricing, Sarah noticed a deeper shift: chip entrepreneurs are getting younger.
In the past, a chip entrepreneur born in 1988 would be considered young. In 2026, she frequently encounters PhDs born in 1995 or 1996 serving as CEOs or co-CEOs of startups.
These young PhDs may not possess over a decade of complete chip mass-production experience, but they can judge technical directions themselves and then hire people with five, ten, or more years of engineering experience to supplement design, verification, and engineering capabilities.
A capital boom changes not just how many companies are founded. It also reprices capabilities within the chip industry and alters who can sit in the top seat earlier.
05、Money Is On The Table, Yet Some Choose Not To Start Up
In mid-2026, Jie Shu, who has spent over a decade in the chip circle, found himself being encouraged to start a business more frequently. People approached him almost every few weeks.
"In March, four groups contacted me, even bringing angel and Series A funds to negotiate," he said. When the market was hottest in May, someone directly told him: "Don't worry about money; we can help you raise all subsequent funds for2026."
His answer was: "You shouldn't start a business when a sector is at its hottest, but rather bet on areas unnoticed by others two or more years in advance."
In Jie Shu's view, if one had focused on this area before 2024, weathering the two years when it was unnoticed, by the time the windfall arrived in 2026 and consensus formed, one would already be a leader in the field. This is a more reliable path, but it requires entrepreneurs to have industry insight and predictive ability.
However, seeing the market heat up by late 2025, founding a company in March 2026, and expecting to raise billions within months makes this path hard to replicate.
For an emerging field, the entrepreneurial window and the financing window are not the same window.
This is what Jie Shu is most wary of in this boom. "Sometimes the development of a story depends not just on the storyteller, but on what kind of story believers need."
A sudden spike in value for a technical direction doesn't necessarily mean entrepreneurs are better at storytelling; it might simply be that capital needs a target capable of carrying a new narrative at that point in time.
Jie Shu chooses not to enter the fray, continuing to accumulate with excellent teams. Others face not the question of "should I start up," but whether what they already have is worth giving up.
Dong Yong is a middle manager at a listed domestic GPU company. In 2026, several groups approached him hoping he would join a startup as a co-founder, but he rejected them all.
"The company's stock options still need a few years to vest. Before they vest, I definitely won't resign."
Beyond options, health is also a concern. Starting a business means enduring high-intensity work and uncertainty again. For a technical manager who has worked in the industry for many years, this calculation is completely different from that of younger people.
The same capital window, for a PhD in his twenties, might mean becoming a CEO ten years earlier.
For a technical lead who has worked at big firms for ten or twenty years, it means abandoning stocks, income, organizational resources, and career security to gamble again on a financing window that might last only half a year.
With CNY 50 million on the table, Dr. Zhang tried to step in, then stepped back out.
Jie Shu did not enter when the sector was booming and capital was most frenzied, but kept a close watch on industrial development directions.
Dong Yong chose to stay at the already-listed company.
The 2026 chip startup boom was hot enough, but it didn't lead everyone to the same conclusion. Are you afraid of missing out, or choosing to miss out? Add the author's WeChat BENSONEIT for more information on startups.
06、Capital Begins To Cool, Chip Talent Prices Still Rising
Han Peng, who has worked in chips for over twenty years and is undertaking his second startup, originally judged that this wave of capital heat would last at least until mid-2027.
But by the second half of 2026, he clearly felt the market cooling down.
For someone who has done chips for over twenty years, a few months is originally not long. But in 2026, a few months are enough to cross an entire capital cycle.
In the first half, startups worried about missing the window of rapid valuation growth. In the second half, worries shifted to whether anyone would still be willing to provide funds two years later.
The listings and high market caps of companies like Moore Threads and MetaX once gave the primary market a renewed view of exit opportunities for chip investments, rapidly amplifying investors' risk appetite. By the second half, as the secondary market weakened, the primary market began recalculating risks.
Spring financing was an offense; by autumn, financing began turning into ammunition reserves.
Yet as capital cooled, chip talent prices continued to rise.
Sarah clearly felt that the primary market was weaker after June than in the first half. Some institutions had already consumed a significant portion of their annual investment quotas in the first half. However, recruitment demands in her hands continued to increase.
In spring, investors injected money into startups; companies used the funds to build teams and initiate hiring. By summer and autumn, this money began flowing centrally into the talent market.
A seemingly contradictory scene appeared in the second half of 2026: chip investment had begun to cool, but chip talent prices were still rising.
Sarah judges that the true inflection point of this hiring boom will be clearer after March or April 2027.
Capital heats up first, and cools down first.
Talent heats up later, and cools down later.
And chip R&D is slower than both capital and talent.
07、Listing Can Be The End For Capital, But Not For Chips
The capital boom gave new architectures like LPUs, 3D, and compute-in-memory a rare opportunity, but capital never comes for technological ideals. It seeks returns and exits.
After the listing of the previous batch of domestic GPU companies, investors were once again willing to put money into chips. For entrepreneurs, listing or acquisition also implies another very realistic thing: personal wealth may undergo massive changes.
A veteran primary market investor noted that some entrepreneurs he had contacted were quite direct: They considered future listing and exit strategies from the moment the company was founded. "They also want to cash out."
This is easily understood. Engineers giving up big-company salaries, stocks, and certainty to start businesses, hoping to exchange higher risk for higher returns, is part of the commercial world.
What is truly alarming is that when returns for both investors and entrepreneurs increasingly depend on the next round of financing and eventual listing, the company's timeline becomes increasingly aligned with the capital timeline.
Product validation hasn't happened, but the next round of financing has already begun.
Customers haven't used it at scale, but valuations can jump several times over.
Industrial investor Huang Zhiyuan, observing this wave of LPU projects, noted that some companies saw valuations surge from approximately CNY 100 million to CNY 2-3 billion. If valuations rise in every round while products, business, and customers don't progress synchronously, it ultimately becomes just a "pure story inflation."
Financing can jump, but industries do not develop by jumping according to financing rhythms.
Listing allows investment institutions to exit and enables entrepreneurs to achieve wealth leaps, but chips ultimately must enter customers' real businesses.
LPUs must prove that token costs are genuinely lower. 3D chips must prove that high-bandwidth advantages are realized in actual scenarios. Customers will not hand over core business to an unverified chip just because a company recently completed a high-valuation financing round.
Wang Xiaofeng and his team increasingly understand how the capital market works.
But the question investors first asked him—what programming model should an LPU chip adopt—has not disappeared because the company's valuation rose.
In the 2026 capital market, two to three months are enough for a hot chip company to complete a financing round.
In chip R&D, two to three months might just be the beginning of solving an engineering problem.
After listing, acquisitions, and wealth realization, the most fundamental question remains: Is anyone actually using this chip?
* Note: Dr. Zhang, Wang Xiaofeng, Li Wei, Chen Junjie, Rocket, Jie Shu, Dong Yong, Han Peng, and Huang Zhiyuan are pseudonyms.